What is ChatGPT Work?
On July 9, 2026, OpenAI introduced a new agent mode for ChatGPT called ChatGPT Work. According to the official announcement from OpenAI, the tool operates in several steps: In a “Plan Mode,” the agent first gathers context from connected tools, asks follow-up questions if necessary, and presents the user with a step-by-step plan for approval. After that, ChatGPT Work continues to work independently – for several hours if needed – accessing files and applications through an integrated browser function on the desktop and through so-called computer-use capabilities.
From the gathered information, the agent creates finished work products: tables, documents, presentations, and, through the new “Sites” feature, also interactive web applications, dashboards, and reports. According to OpenAI’s product page, one-time or recurring tasks can also be set up, with progress tracking available. Integration with external tools is facilitated through more than 1,400 plugins, including for Slack, Salesforce, Google Drive, SharePoint, Jira, and common email programs. A security mechanism is said to prevent unauthorized data sharing, according to OpenAI; enterprise customers can centrally manage access rights and connected tools.
GPT-5.6: Sol, Terra, and Luna in Price Comparison
Alongside ChatGPT Work, the new model family GPT-5.6 is publicly available. It consists of three tiers, which differ in price per million tokens according to OpenAI’s pricing overview in the Help Center:
- Sol (flagship): $5 input / $30 output
- Terra (mid-tier): $2.50 input / $15 output
- Luna (cheapest, fastest tier): $1 input / $6 output
Sol was previously available as a limited preview for selected partners, as OpenAI describes in the blog post “Previewing GPT-5.6 Sol”. There, the company mentions record scores on the coding benchmark Terminal-Bench 2.1, as well as improved performance compared to GPT-5.5 with lower token consumption on the biology benchmark GeneBench v1. On the cybersecurity benchmark ExploitBench, Sol reportedly achieves a comparable result to the internal model “Mythos Preview” at about one-third of the output tokens. According to OpenAI, Sol does not exceed the threshold classified as “Cyber Critical” in the company’s own Preparedness Framework. These statements come from OpenAI’s own publications and have not yet been verified by independent third parties.
Rollout and Availability
The rollout occurs in stages based on the plan. As OpenAI states in the official announcement, ChatGPT Work has been available since July 9, 2026, initially to Pro, Enterprise, and Edu users via web and mobile. Plus and Business plans are expected to follow within a few days. The revised desktop app for Mac and Windows – which also integrates the previous Codex app – is already available today for all plans, including the free version. The previous ChatGPT desktop app will be renamed “ChatGPT Classic,” according to MacRumors and will remain available in parallel.
Competitive Comparison and Initial Reactions
ChatGPT Work directly competes with existing agent products from other providers. MacRumors explicitly places the tool alongside Anthropic’s Cowork in its reporting, which also operates in the background across multiple applications. OpenAI also positions the GPT-5.6 model family against current competing models like Anthropic’s Fable 5.
Initial assessments from testers, compiled by Axios, vary. MagicPath CEO Pietro Schirano described GPT-5.6 as “the best model I’ve ever used. Fast, smart, genuinely creative.” Investor Matt Shumer, on the other hand, was more reserved, noting that Anthropic’s Fable was “quite a bit better” for his use cases. Every CEO Dan Shipper compared GPT-5.6 to a reliable but unremarkable tool, while attributing higher raw performance to Fable. According to Axios, OpenAI CEO Sam Altman also quantified the efficiency increase of Sol in agent-based coding tasks compared to predecessor models at 54 percent – a company-specific statement that has not yet been independently verified.
Classification: What the Numbers Are Worth
Part of the previous reporting relies on material published by OpenAI itself, such as benchmark results and a customer example from Virgin Atlantic cited in the official announcement, stating that a competitive analysis cycle that normally takes weeks now “takes hours.” Such company-specific case studies and efficiency claims are naturally not independently verified and should be accordingly contextualized. How robust the advantages of ChatGPT Work and GPT-5.6 are in everyday use will likely only become clear when the rollout is expanded to broader user groups and independent tests are available.
